(Source from Reuters/Alamy)Defining the meaning of image recognition: Importing specific information from image data.
– At this point, image recognition has developed by recognizing the type and location of the object, and the pattern of image data pixel values. Recently, the perception of the object is performing beyond the human level, and various objects and their locations are being studied.Here we learn the basic way to get the type and location data of the object. To find out the above, the most basic way is to identify patterns using filters.
In the field of imaging, even if a filter is considered to be a material for extracting the “similarity” of two images through convolutional operations, the “similarity” of the filter size part is extracted by multiplying the value of the filter by the value of the input image.
If this value is large, it means that the filter and the image are the same, and if it is small, it means that it is not the same, so it can be scaled with probability using a function such as sigmoid
The concept of CNN, a deep learning model, is as above.
A specific filter is compounded into the image, and strides one space at a time.
ConV -> pooling based on the fact that if similar data exists in the space pointed by the filter, the value increases or the value is low
You can extract data by doing.Due to the above process, similarity verification using filters is the main method of image recognition, and among deep learning architectures, there are models that only perform this part of the operation and remove the fully-connected layer.
– You can learn the filter, too.
You can filter representative images for any class.
(Source from Reuters/Alamy)There is also a way to organize and create a system that can be used, i.e., run Conv with various filters and sizes for the entire image, and then select the high-value section, class classification and location navigation are completed.Class and size issues.
Even if it’s the same class, it can have different shapes.
Even if the average value is used, it depends on the angle, so it is necessary to find a method to have a filter for it.If there are N samples, N operations per pixel are required.
Also, size is a problem.Since the size of the image and the size assumed by the pixel are different, the size of the filter is also changed to find the object in it
If you change the size M times, you need another M operations per pixel.
Eventually, the operation NxM per object must be applied to all image data pixels of size WxH, and if there are multiple object classes, the above operation must be repeated as many times as possible. As explained above, image recognition is complicated and diverse. And it’s very important to apply image recognition technology to how and where. These companies are GB SOFT (CEO: Ki-bum Park). This company approaches various fields through artificial intelligence solutions through videos and measures, manages, and develops DBs of general consumers and corporate customers. First of all, it is applied to the safety management of workers in the construction civil engineering field and the access and security of large-scale apartments to deliver and manage software to large construction companies. In addition, it was applied to large-scale production lines as a program to manage the status of workers. Various accessibility is considered an advantage of development companies, so attention is being paid to what kind of moves they will take in the future
▲ GB Soft Co., Ltd.
▲ CEO : Ki-bum Park
▲ http://gb-soft.co.kr
▲ gbsoft@gb-soft.co.kr
▲ +82-53-716-4884
HS HA
ASIA JOURNAL
